Cross-domain few-shot learning with unlabelled data

Abstract

Few shot learning aims to solve the data scarcity problem. If there is a domain shift between the test set and the training set, their performance will decrease a lot. This setting is called Cross-domain few-shot learning. However, this is very challenging because the target domain is unseen during training. Thus we propose a new setting some unlabelled data from the target domain is provided, which can bridge the gap between the source domain and the target domain. A benchmark for this setting is constructed using DomainNet \cite{peng2018oment}. We come up with a self-supervised learning method to fully utilize the knowledge in the labeled training set and the unlabelled set. Extensive experiments show that our methods outperforms several baseline methods by a large margin. We also carefully design an episodic training pipeline which yields a significant performance boost.

References in corpus (10)

• A Simple Framework for Contrastive Learning of Visual Representations
• Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
• Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks
• Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID
• A Closer Look at Few-shot Classification
• Moment Matching for Multi-Source Domain Adaptation
• Feature Alignment and Restoration for Domain Generalization and Adaptation
• A Broader Study of Cross-Domain Few-Shot Learning
• Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles
• Disjoint Label Space Transfer Learning with Common Factorised Space

Cited by in corpus (1)

• TGDM: Target Guided Dynamic Mixup for Cross-Domain Few-Shot Learning